PRAGMA — Revolut’s Finance Foundation Model

Summary: PRAGMA is Revolut’s proprietary foundation model trained on 24 billion banking events, delivering major uplifts across credit scoring, fraud, and marketing vs. prior ML models.

Sources: raw/articles/simon-taylor-2026-04-26.md, raw/call-notes/carlos-2026-05-10.md

Last updated: 2026-05-17


What It Is

PRAGMA is a behavioral foundation model — not an LLM. It doesn’t generate text. It reads sequences of customer events (logins, screen taps, payments) and learns rich representations of customer behavior.

  • 26M customers, 24B customer events, 207B tokens
  • Trained on 32 H100s in ~2 weeks
  • Built with NVIDIA (NeMo AutoModel, cuDF for feature engineering)
  • Replaces six separate custom ML models with one

Training Approach

Three experiments to validate the model:

  1. Pre-trained embeddings only — how much information is in the embedding before task-specific training?
  2. Embeddings + hand-crafted ML features — what does the foundation model capture that years of feature engineering missed?
  3. LoRA fine-tuning — can the foundation model beat the entire data science team by pressing a button? (Answer: mostly yes.)

Production Uplifts vs. Prior ML Models

Use CaseUplift
Credit scoring (PR-AUC)+130%
Fraud recall+65%
Marketing engagement+79%
Product recommendation+40%
Anti-money laundering−47% (expected failure — AML is a network problem; PRAGMA reads users in isolation)

Why AML Failed

AML is a network problem — what matters is who you transact with, not what you do. PRAGMA reads each user’s history in isolation and can’t see transaction chains. This is a known limitation; the team is working on the next architecture to address it.

What’s Next: Generative Finance Models

PRAGMA today is like BERT in 2020 — it reads and predicts. The next step is generative: a model that could write a customer’s future event sequence, simulating when they’ll take a product, then rewinding to identify what conditions led there and manufacturing those conditions.

Business Case

Napkin math on established institutions:

  • JPMorgan has $10B+ annual credit costs. Even 10% of PRAGMA’s stated credit scoring gain → hundreds of millions/year.
  • Every 5.75 in operational overhead. A 65% recall improvement pays for the GPU bill many times over.